Launchpad bug pages trigger caret browsing in Firefox and other Gecko browsers

LP: https://bugs.launchpad.net/bugs/107247

Steps to reproduce:
1. In Firefox or Epiphany, open any launchpad bug page, such as the one above.
2. Wait for the Subscribers box to finish loading.
3. Press the Up or Down arrow keys.

Result: firefox switches to caret browsing mode.

We see this on ffox 3.0/3.5 and mozilla-central builds.

Looks like the linked site changed something since this bug was filed so that it no longer qualifies as a test case. Does anyone have another test case?

hmmm. i think you need to be member of the launchpad beta testers to see the version affected. i will try to get something ...

got more info: seems it should also happen if you are not in the launchpad beta testers. however, i was told that you need to be logged in to reproduce.

maybe also helpful that this commit makes it goes away:

http://bazaar.launchpad.net/~launchpad-pqm/launchpad/devel/revision/9000

Sorry, but there's no obvious link between that commit and caret browsing.

maybe also check https://bugs.edge.launchpad.net/bugs/404088 (and https://bugs.edge.launchpad.net/bugs/392138) which was the bug that caused the above fix for this as a side-effect to land. They are talking about focusing some hidden element. does that ring a bell?

No, but I'll CC enndeakin as being the person who knows most about focusing...

> Result: firefox switches to caret browsing mode.

That seems unlikely. Do you actually mean that caret browsing is enabled (and that the preference is set), or just that the caret appears? Is the text editable?

Please describe what actually happens.

A testcase is needed here as well.

we are working on a testcase ...

But it really triggers caret browsing. I just tried and this are the instructions

 1. caret browsing disabled (accessibility.browsewithcaret=false)
 2. log into launchpad and go to the bug page and wait till the subscribers are loaded
 3. press the up and/or down key

(
Expected result: page scrolls
Actual result: page doesn't scroll as expected - feels like a missing focus.
)
 4. click on the page (i clicked in the title)
 5. use arrow keys

Expected result: page scrolls
Actual result: a caret appears which you can move on the document using the arrow keys.

 however, a caret appears and the arrow key move it through the document (which i refer to as caret browsing).

> 4. click on the page (i clicked in the title)
> 5. use arrow keys

Also, sometimes you dont need those two extra steps

(In reply to comment #8)
> that the preference is set), or just that the caret appears? Is the text
> editable?

no its not editable. you can just move the caret through the document using arrow keys.

This is still happening today (2009-08-07) on the main (deployed) Malone instance, eg. the URL:

  https://bugs.launchpad.net/malone/+bug/392138

exhibits the behaviour on about 50% of page loads.

Yes, just to confirm, the page *is* triggering the caret to appear in between the characters in the text, and that:

  accessibility.browsewithcaret=false

I reproduced this first time, by opening the above URL, PageDn, PageDn, PageDn (until at bottom), PageUp, PageUp, PageUp (until at top), and which point the PageUp/Down stopped working and the caret had appeared, movable via the Up/Down arrow keys.

Repeated tests to try to reproduce it (each started from a fresh page load):

1. Yes: PageDn (x3), PageUp (x3), page "locked", caret there
2. Yes: PageDn (hold), PageUp (hold), page locked, click for caret to appear
3. No.
4. Yes: PageUp (repeat), PageDn (repeat), PageUp (repeat), PageDown (repeat), Arrow Down, caret on
5. Yes: Hold Down Arrow whilst page loading, Page jumps back (almost) to top, page "locked", click for caret to show
6. Yes: Hold Down Arrow whilst page loading, scroll down, jump back to top, scolls down again, jump to top again, page locks; click to different tag, click back again, page unlocks.
7. Yes: Page locked itself, click for caret to appear, change tabs and back again, page unlocked.
8. Yes: Hold Arrow Down whilst page loading, page jumped + locked, change tabs and back to unlock.
9. Yes: Hold Up Arrow whilst page loading, page locked
10. Yes: Hold Page Down whilst page loading, page jumped to top and locked, click for caret to appear.

So, nine times out of ten, reproducibility. The most reliable way (although not the only way) appears to be to hold down one of the scrolling (Arrow or Paging) keys whilst the page is loading. And that changes tabs clears caret browsing mode.

The Downstream bug report has that produce cross-platform (and cross browser) so appears to be Gecko itself kicking into caret mode, rather than the browser interface enabling it.

It appears that bug #289632 ("Caret browsing halts keyboard focus on javascript links") may be related to this. There are certainly <a href="..." onclick="..."> elements in the Launchpad page in question.

Still can't reproduce this.

Thanks Neil,

This is still 100% reproducible by my end, simply by holding the Down Arrow key down whilst loading the page of any Launchpad Bug report. *Whilst logged in*.

The being *logged in* bit is crucial, as otherwise the box that is triggering this caret behaviour will not be made editable.

Is there anything further that can myself or other can do to assist?

  -Paul

http://launchpadlibrarian.net/30122852/prevent-scrolling.html is an attachment to the downstream bug which exhibits a similar problem, although it's not 100% the same as the Launchpad issue. Apparently Paul was unable to isolate the precise HTML/CSS/JavaScript combo which would allow you to reproduce fully, but his attachment should at least show you that something is wrong, and that this is reproducible. (See also the comments at https://bugs.launchpad.net/malone/+bug/107247/comments/28)

I've been noticing this same behavior on linux when navigating through google search results with the keyboard. Pressing PgUp/PgDn sometimes enables caret browsing; the caret shows up (sometimes not immediately visible) and PgUp/PgDn stop functioning (movement of the caret with the cursor keys still works). Pressing F7 and selecting "No" restores PgUp/PgDn behavior to normal.

The caret seems to only be enabled for the tab with the search results; other tabs are unaffected. Switching tabs or windows and coming back to the search results is not sufficient to turn off caret browsing for that tab.

The problem is intermittent; I have not yet been able to reproduce it at will.

I also have occasionally had this issue. Happened again just today after a quick Google search, page down was no longer working and I discovered caret browsing had been enabled.

Pressing F7 results in the prompt box ("Pressing F7 ... Do you want to turn caret browsing on?") and I press no which causes it to disable and allows page up/down and the arrow keys to work once again.

I tried repeating my google search but it did not enable again the second time. I've not been able to reproduce it with any consistency but it does happen occasionally.

Split from bug 549262.

1. Load the testcase, https://bugzilla.mozilla.org/attachment.cgi?id=459420
2. Press the down arrow key several times.

Result: scrolls to the top of the pink div and stops.

Happens on Windows too => OS = All.

Possibly even a dupe of bug 549262...

(In reply to comment #2)
> Possibly even a dupe of bug 549262...

Oh, forget this comment, I was totally confused.

Boris, is it possible to specify a "fallback" action for the arrow keys?

I'm on fire today! Boris, see comment 3. :-)

I don't know offhand.... Neil might.

(In reply to comment #5)
> I don't know offhand.... Neil might.
I don't understand the question...

So, here is the problem. In platformHTMLBindings.xml, we define a VK_DOWN handler for caret navigation command so that you can go to the next line in editor by pressing down for example. If there is no editable element focused, we need the default behavior of scrolling the page down to take place, so we need some kind of a fallback event handler being bound to VK_DOWN. Is there any way to do that?

(In reply to comment #7)
> So, here is the problem. In platformHTMLBindings.xml, we define a VK_DOWN
> handler for caret navigation command so that you can go to the next line in
> editor by pressing down for example. If there is no editable element
> focused, we need the default behavior of scrolling the page down to take
> place, so we need some kind of a fallback event handler being bound to
> VK_DOWN. Is there any way to do that?
There is a way to do that, but you have to bind your handlers and attach your controllers to the focused element rather than the window root as you currently do.

The second approach is to fix the editor's command controller to work with editable regions.

The third approach is to stop editor from handling the up and down arrow keys. If you look at the global window controller you'll see that if it thinks there's a caret then it will try to do caret movement instead of scrolling. I think the test detects design mode but not contentEditable regions, so it still tries to do scrolling in that case.

I just encountered this issue myself, on Planet Mozilla (planet.mozilla.org). I visited Planet Mozilla, and noticed that I couldn't seem to use the arrow keys to scroll. Poking at it further, I found that if I held down an arrow for a while, the page would start scrolling in that direction after a bit; if I stopped, and then pressed in the same direction, it would scroll in that direction immediately, but reversing direction required waiting a bit. This felt a lot like caret browsing; however, caret browsing was not actually turned on. I pressed F7 to turn on caret browsing, and after confirming that in the dialog, the caret appeared exactly where the scrolling behavior suggested it would. (For example, if I pressed down until the page started scrolling, then pressed up a couple of times, then turned on caret browsing, the caret would appear a couple of lines above the bottom of the window.)

Build identifier: Mozilla/5.0 (X11; Linux x86_64; rv:7.0a2) Gecko/20110715 Firefox/7.0a2

Just in case something changes, I'll attached a snapshot of Planet Mozilla. I've confirmed that the problem occurs with only the HTML present, without the CSS, scripts, images, or videos.

Created attachment 548844
Snapshot of Planet Mozilla from 2011-07-27, which triggers invisible Caret Browsing

And I can confirm that clicking on the attachment and browsing it here on Bugzilla still triggers the same behavior.

Likely caused by the post that uses contenteditable.

(In reply to comment #23)
> Likely caused by the post that uses contenteditable.

Indeed.

*** This bug has been marked as a duplicate of bug 669026 ***

*** Bug 505732 has been marked as a duplicate of this bug. ***

I also noted that middle-clicking links on Planet Mozilla doesn't work.
Same bug?

(In reply to comment #8)
> The third approach is to stop editor from handling the up and down arrow
> keys. If you look at the global window controller you'll see that if it
> thinks there's a caret then it will try to do caret movement instead of
> scrolling. I think the test detects design mode but not contentEditable
> regions, so it still tries to do scrolling in that case.
It does detect contentEditable regions. But some focus code to make link focusing work in caret browsing mode blurs the region :-( See nsSelectMoveScrollCommand::DoCommandBrowseWithCaretOn()

Created attachment 548924
Patch (v1)

*** Bug 481626 has been marked as a duplicate of this bug. ***

Created attachment 548949
Proof of concept

The editor-base.inc change remaps VK_UP to cmd_scrollLineUp because that was easier to prototype, but ideally it would be renamed back to cmd_linePrevious throughout the codebase. The editor command handler is ruled out of the equation, and that code should be removed. The nsGlobalWindowCommands already knows how to move a caret, so that doesn't change, but the focus manager has to be told not to blur a contenteditable region.

Note that even with caret browsing on, this code doesn't move focus in to a contenteditable region either. I don't know whether this is reasonable.

I don't understand why the focus manager change there is OK...

Why don't you make nsEditor::IsActiveInDOMWindow() an insterface method of nsIEditor?

And, PageUp, PageDown, Home, End keys need to be handled too. (I'm not sure about Shift+Arrow or them)

(In reply to comment #25)
> I also noted that middle-clicking links on Planet Mozilla doesn't work.
> Same bug?

No, please file one?

mbrubeck filed it: bug 674770

*** Bug 674860 has been marked as a duplicate of this bug. ***

Comment on attachment 548949
Proof of concept

Well, I might as well ask for review on the nsFocusManager change, just in case... the point here is that we don't normally get here because the editor command handlers are active in contenteditable mode, but I want to be able to remove the editor motion command handlers because the global ones work anyway, except for the issue that they blur contenteditable regions, which is unwanted.

Ehsan, does Neil's approach sound better to you? It does to me.

Is this a regression in 6? How widespread is this likely to be? We're just a few days from the final Beta of 6.

(In reply to comment #19)
> Is this a regression in 6? How widespread is this likely to be? We're just a
> few days from the final Beta of 6.

This is not a regression in 6, from what Ehsan said on IRC, but e.g. planet.m.o is unusable using keys for several days now, due to this bug.

(In reply to comment #20)
> (In reply to comment #19)
> > Is this a regression in 6? How widespread is this likely to be? We're just a
> > few days from the final Beta of 6.
>
> This is not a regression in 6, from what Ehsan said on IRC, but e.g.
> planet.m.o is unusable using keys for several days now, due to this bug.

While that is true, this bug has existed for ages (I think from Firefox 3+) and there is no good reason why we should track it for Firefox 6...

(In reply to comment #18)
> Ehsan, does Neil's approach sound better to you? It does to me.

As I've said before, I don't understand the focus manager code, so if Neil (Deakin) vouches on that, I'd be fine with Neil's approach. I'm assuming that he's going to address the todo items in comment 13 once Neil reviews the focus manager change. Also, Neil, can you check to see if my test case passes successfully with your patch?

(In reply to comment #13)
> Note that even with caret browsing on, this code doesn't move focus in to a
> contenteditable region either. I don't know whether this is reasonable.

That's definitely not desired, but I tested it and we are doing the same thing right now even without this patch, so I don't think we need to worry about this problem here.

Comment on attachment 548949
Proof of concept

The focus manager change looks ok, but the editor-base.inc change is not.

The former fixes an issue where there is a selection within the editable area, but the editable area is not focusable, then tab is pressed. (The removed code from nsFocusManager is only called when there is nothing focused).

The editor-base.inc change though makes scrolling occur in a number of situations where it shouldn't, for example, scrolling should not occur while caret browsing is on, nor while the caret is within the focused editable area.

Is the editor attaching it's command controller to the document and not the editable area?

(In reply to comment #23)
> The former fixes an issue where there is a selection within the editable
> area, but the editable area is not focusable, then tab is pressed. (The
> removed code from nsFocusManager is only called when there is nothing
> focused).
I could only get that code to be hit when there was a caret. I'm not sure how to make an unfocusable editable area, but I guess it won't have a caret.

> The editor-base.inc change though makes scrolling occur in a number of
> situations where it shouldn't, for example, scrolling should not occur while
> caret browsing is on, nor while the caret is within the focused editable
> area.
The global window commands already know about caret browsing. (If you look at the browser-base.inc then you'll see the bindings are the same. At some point we may want to rename them back to reduce confusion.) And, except for the code block in the focus manager, its idea of a caret also applies to editable areas, which was why I wanted to remove the code block!

> Is the editor attaching it's command controller to the document and not the
> editable area?
Correct. There was some issue with attaching the commands to the editable area which ehsan has pointed me towards several times but I still forget the bug #.

(In reply to <email address hidden> from comment #24)
> > Is the editor attaching it's command controller to the document and not the
> > editable area?
> Correct. There was some issue with attaching the commands to the editable
> area which ehsan has pointed me towards several times but I still forget the
> bug #.

I tried very hard in bug 289384 to do that, but I failed miserably... Maybe because I didn't have a good idea what I'm doing. :/

Comment on attachment 548949
Proof of concept

I'm not sure if something else is being asked of me here.

With this patch:
 1. Click in the editable area
 2. Press cursor up/down

Expected:
  The caret moves up/down a line
Actual:
  Page scrolls, so keyboard users cannot edit other lines.

(In reply to Neil Deakin from comment #26)
> Expected:
> The caret moves up/down a line
Well, that's what I get locally. Does caret browsing work in your build?

I get the same whether caret browsing is enabled or not.

(In reply to Neil Deakin from comment #28)
> (In reply to <email address hidden> from comment #27)
> > (In reply to Neil Deakin from comment #26)
> > > Expected:
> > > The caret moves up/down a line
> > Well, that's what I get locally. Does caret browsing work in your build?
> I get the same whether caret browsing is enabled or not.
Sorry, I wasn't clear, I meant caret browsing in general, not just in editable areas.

Ping?

*** Bug 681418 has been marked as a duplicate of this bug. ***

just in case, HOME/END key is breaks too...

There seems to be some expectation that I respond here but I have nothing else to add. The patch creates an issue as described in comment 26. The same issue occurs with caret browsing enabled or disabled.

Caret browsing doesn't work with this patch. The page scrolls when I try to move the caret up and down.

Created attachment 565950
Fixed proof of concept

Comment on attachment 565950
Fixed proof of concept

Testing and debugging shows that this seems to work.

Neil, does your patch make the test part of attachment 548924 pass? If so, we can get them landed. :-)

(In reply to Ehsan Akhgari from comment #37)
> If so, we can get them landed. :-)

Except it's very much proof of concept (although I wouldn't mind landing the nsFocusManager.cpp change which would be necessary anyway).

In particular, I'm not sure a) which keys need to be changed b) whether those keys should map to the same commands in textareas and browsers/editors c) how native Linux key bindings fit in.

Created attachment 569554
Proposed patch

The names of the commands are configured in three places:
* GTK2 Native keybindings
* Editor bindings (editor-base.inc and libeditor)
* Browser bindings (browser-base.inc and nsGlobalWindowCommands)
The browser bindings have six differences to the other two sets of bindings,
all cmd_scroll* names, which I have renamed for consistency.
The browser bindings were missing implementations for selecting by page.
The stubs also used the wrong command names, so I corrected that too.
This then allowed me to remove the editor implementation.

Comment on attachment 569554
Proposed patch

Review of attachment 569554:
-----------------------------------------------------------------

Looks very good!

Does my test case pass with this patch?

(In reply to Ehsan Akhgari from comment #40)
> Does my test case pass with this patch?
Yes, it does.

So, things didn't go quite as well as I'd hoped.

Text areas and input fields need to have the editor behaviour.
Design mode and contenteditable windows need to have the global behaviour.

Created attachment 570250
Fixed patch

So, I ended up having to separate the cursor movement commands into their own command table so that I could append that controller for textareas and inputs.

Neil, any news about this bug?
It affects Fx9+(current beta), if we don't fix it in Fx11, then user have to wait at least 3*6=18 weeks before thay can copy text in those page again...

Sorry, I mistaken this bug for bug 692153.

(In reply to dindog from comment #44)
> Neil, any news about this bug?
ehsan, should I be asking someone else for review?

> It affects Fx9+(current beta), if we don't fix it in Fx11, then user have to
> wait at least 3*6=18 weeks before thay can copy text in those page again...

(In reply to dindog from comment #45)
> Sorry, I mistaken this bug for bug 692153.
Actually it turned out to be fairly straightforward to adapt this patch to apply to Copy, but it won't help with the Select All case.

Comment on attachment 570250
Fixed patch

Review of attachment 570250:
-----------------------------------------------------------------

I'm not crazy about "editor" vs. "editing" controllers, but I couldn't really think of anything better, so r=me.

Thanks a lot for your work on this, Neil, and sorry that I did a very bad job at responding to the review request in time.

Try run for e3859a3ea446 is complete.
Detailed breakdown of the results available here:
    https://tbpl.mozilla.org/?tree=Try&rev=e3859a3ea446
Results (out of 208 total builds):
    success: 144
    warnings: 50
    failure: 14
Builds (or logs if builds failed) available at:
http://<email address hidden>

test_showcaret.xul: uses cmd_scrollBottom, got renamed to cmd_moveBottom
test_selection_move_commands.xul: many renames...
test_movement_by_characters/word.html: rely on the buggy behaviour!

test_selection_move_commands actually tests for the broken commands. Sigh...

test_movement_by_characters.html's reliance on this bug is due to it failing to focus the content before performing the navigation.

test_movement_by_words.html has the above bug but did also find a real bug; with this patch, you can't move by words to the very end of a contenteditable area, because it tries to move right out of the contenteditable area...

Try run for 8c5b9300caf0 is complete.
Detailed breakdown of the results available here:
    https://tbpl.mozilla.org/?tree=Try&rev=8c5b9300caf0
Results (out of 208 total builds):
    exception: 1
    success: 137
    warnings: 56
    failure: 14
Builds (or logs if builds failed) available at:
http://<email address hidden>

Created attachment 584062
With test fixes (for check in)

Try run for f0084cbf3b9d is complete.
Detailed breakdown of the results available here:
    https://tbpl.mozilla.org/?tree=Try&rev=f0084cbf3b9d
Results (out of 206 total builds):
    success: 167
    warnings: 24
    failure: 15
Builds (or logs if builds failed) available at:
http://<email address hidden>

> warnings: 24
> failure: 15
Most of these are hidden; 3 randomorange show up on tbpl by default.
More oranges and the reds show up with noignore or via self-serve but they're all Lion, Tegra and/or Jetpack so nobody cares about them.

This bug is fixed in 20121224 nightly. Neil, would you please look into the copy menuitem too?

Ehsan, sorry again for the other day urge you to review this bug in IRC. I got a reason to worried. Two major sites in China use contenteditable element in part of there service, (In fact, they rank No.1 & No.2 in China and No.5 & No.10 all over the world in Alexa), both are popular things, try imagining as they are Chinese Facebook and Google group, you'll got the picture.

I heard complaint about unable to copy and scroll after Fx9.0 was released. I don't want Firefox losing users. I'll would be better have it fixed not only in nightly, but also beta, even release.

*It* 'll would be better have it fixed not only in nightly, but also beta, even release.

sorry for the misspell

Pushed changeset 4d0391866459 to mozilla-central.

(In reply to dindog from comment #56)
> Neil, would you please look into the copy menuitem too?
Yes, I'll do copy (but not select all, because it's harder) in bug 692153.

"Dude, where's my commands?!" This change removed commands that have been around for many years, breaking add-ons: bug 714164.

Neil, I think you need to add those commands back, and just make us not use them... :(

Fair enough, and I can put the test back too, but I might as well use bug 714164.

*** Bug 702064 has been marked as a duplicate of this bug. ***

Bug 696020 appears to have exacerbated this problem in 10 (see bug 702064). The attached patch appears to be higher risk than we're comfortable for Beta (please correct me if I'm wrong), but I'm wondering if this is a good candidate for Aurora 11.

(In reply to Alex Keybl [:akeybl] from comment #63)
> Bug 696020 appears to have exacerbated this problem in 10 (see bug 702064).
> The attached patch appears to be higher risk than we're comfortable for Beta
> (please correct me if I'm wrong), but I'm wondering if this is a good
> candidate for Aurora 11.

I'd be much more comfortable if we could back out bug 696020 from Aurora. smaug, can we do that?

No. We were thinking to release 9.0.2 with bug 696020 fixed.

Or if we do that, we need to backout all
bug 696020, bug 689564 and bug 659350.

*** Bug 731924 has been marked as a duplicate of this bug. ***

If I could request a summary of the status, where are we on this bug now? dindog@163's saying that it's fixed in 20121224 nightly? How about in release? When is this going to be seen in a release, and when?

If I've done my sums right, you should find that it's currently available in Firefox Aurora, will become available in Beta on the 13th, and then released on the 24th of April. (If you think 17 weeks is a long time to wait, don't forget that Firefox 4 came out 14 months after Firefox 3.6!)

Woot, I see this now in Firefox 12. (This is the main reason we're still using a designMode iframe in G+)

(In reply to garryb from comment #70)
> Woot, I see this now in Firefox 12. (This is the main reason we're still
> using a designMode iframe in G+)

Great! Please file new bugs if you see more problems blocking you from moving to contenteditable elements.

As of 13.0.1 this still hasn't been fixed. Please see attached example:
http://pastebin.com/WTiitrQy

(In reply to Jakub Ondrusek from comment #72)

Please file a new Report in the Core/Editor Component with your Testcase attached. Thanks!